Bearing intelligent fault diagnosis method and system based on generative adversarial network data enhancement and deep transfer learning

Abstract

The invention provides a bearing intelligent fault diagnosis method and system based on generative adversarial network data enhancement and deep transfer learning, and the method comprises the steps: enhancing bearing vibration signal data through a generative adversarial network, and extracting deep time-frequency features through deep learning; a deep migration learning strategy is adopted, features are migrated from a source domain to a target domain, and domain self-adaption and parameter migration technologies are included. And based on the migrated features, using a classifier to identify the bearing fault type. According to the method, the combined vibration signal is generated through the auxiliary classification generative adversarial network, so that the problems of high data acquisition cost and sample imbalance are effectively solved, especially the lack of composite fault mode data. A deep learning model is combined with a residual attention mechanism to extract deep features, and a deep transfer learning strategy minimizes inter-domain differences and improves generalization ability. The designed classifier comprises a feature mapping layer, a dynamic weight distribution layer and a decision output layer, and the decision boundary is optimized through a gating mechanism, so that the fault recognition accuracy is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of bearing fault diagnosis, and in particular to a bearing intelligent fault diagnosis method and system based on generative adversarial network data enhancement and deep transfer learning. Background Art

[0002] In industrial rotating machinery, bearings are core transmission components, and their operating status directly affects the safety and reliability of the equipment. According to statistics, approximately 30% of rotating machinery failures are caused by bearing damage. Early, weak fault characteristics are easily drowned out by noise, resulting in severe challenges for traditional fault diagnosis methods. Existing intelligent bearing diagnosis technology mainly relies on deep learning models to analyze vibration signals, but its accuracy is limited. Specifically, the cost of collecting bearing failure data throughout its life cycle is high, and the number of normal and fault samples in actual scenarios is extremely unbalanced, especially the extreme lack of data on complex fault modes. Traditional data enhancement methods (such as superimposed noise and time domain scaling) can only achieve limited dimensional expansion and cannot generate synthetic samples that conform to the actual physical distribution, resulting in a sharp drop in the model's diagnostic accuracy under few-sample or zero-sample fault modes. Summary of the Invention

[0003] The present invention aims to at least solve the technical problem of low fault diagnosis accuracy in the prior art, and particularly innovatively proposes a bearing intelligent fault diagnosis method and system based on generative adversarial network data enhancement and deep transfer learning.

[0004] In order to achieve the above-mentioned object of the present invention, the present invention provides a bearing intelligent fault diagnosis method based on generative adversarial network data enhancement and deep transfer learning, the method comprising: S1. Data enhancement of bearing vibration signals is performed through generative adversarial networks to generate synthetic vibration signals consistent with the distribution of real fault samples. S2. Using a deep learning model to extract features from the enhanced vibration signal to obtain deep time-frequency features; S3. Based on a deep transfer learning strategy, the deep time-frequency features are transferred from the source domain to the target domain to obtain the transferred deep time-frequency features. The deep transfer learning strategy includes domain adaptation and parameter transfer. The domain adaptation achieves knowledge transfer by minimizing the difference in feature distribution between the source domain and the target domain. The parameter transfer adapts to the new domain data by freezing the underlying parameters of the deep learning model and fine-tuning the high-level parameters. S4. Using a classifier to identify the bearing fault type based on the shifted deep time-frequency features, and obtain a fault diagnosis result.

[0005] As an optional embodiment of the present invention, optionally, in step S1, performing data enhancement on the bearing vibration signal by generating an adversarial network includes: S101. Construct an auxiliary classification generative adversarial network, wherein the auxiliary classification generative adversarial network includes a generator and a discriminator. The generator is a one-dimensional convolutional network structure, including multiple upsampling blocks, each of which is composed of a transposed convolution layer, a batch normalization layer, and an activation function, and is used to map the input random noise and category label into a pseudo vibration signal; the discriminator adopts a one-dimensional convolutional network structure, including multiple downsampling blocks, each of which is composed of a convolution layer, a batch normalization layer, and an activation function, and the output of the discriminator includes two branches, one branch is used to output the probability that the signal is true, and the other branch is used to output the category probability distribution of the signal; S102, training the auxiliary classification generative adversarial network using loss function optimization; S103: Compare the pseudo vibration signal generated by the trained auxiliary classification generative adversarial network with the real fault sample, and continuously adjust the parameters of the generator and the discriminator so that the generated pseudo vibration signal is close to the real fault sample in distribution.

[0006] As an optional embodiment of the present invention, optionally, in step S3, based on a deep transfer learning strategy, the deep time-frequency features are transferred from the source domain to the target domain, and obtaining the transferred deep time-frequency features includes: S301, constructing feature spaces of the source domain and the target domain, and using a domain adaptation method to make the feature distributions of the source domain and the target domain consistent by minimizing the KL divergence or the maximum average difference between the feature distributions of the source domain and the target domain; S302: Freeze the underlying parameters of the deep learning model, retain the common features learned in the source domain, and fine-tune the high-level parameters of the deep learning model to adapt to the characteristics of the target domain data; S303: Continuously adjust the parameters of the deep learning model by iteratively optimizing the deep learning model until the migrated deep time-frequency features are obtained.

[0007] As an optional embodiment of the present invention, optionally, in step S4, identifying the bearing fault type using a classifier based on the shifted deep time-frequency features to obtain a fault diagnosis result includes: S401. Construct a classifier, which includes a feature mapping layer, a dynamic weight allocation layer, and a decision output layer. The feature mapping layer adopts a fully connected structure to map the migrated deep time-frequency features to a latent space; the dynamic weight allocation layer generates channel weights through a gating mechanism; and the decision output layer predicts fault probability based on weighted features. S402, training the classifier using a cross entropy loss function, and continuously adjusting the parameters of the classifier to minimize the difference between the true label and the predicted label; S403. During the training process, an early stopping strategy is adopted. When the loss on the validation set no longer decreases, the training is stopped and the optimal classifier is saved. S404: Input the vibration signal of the test set into the trained classifier and output the fault diagnosis result.

[0008] As an optional embodiment of the present invention, optionally, the expression of the classifier is: ; ; ; in, represents the failure probability vector output by the classifier, Represents the deep time-frequency features after the fault classifier is transferred Mapped to a failure probability vector The composite operation, represents the deep time-frequency features after migration, represents the set of learnable parameters of the classifier, represents the Softmax function, represents the projection matrix, represents the graph attention weight vector, represents the fully connected projection matrix, represents the bias vector, represents the Sigmoid activation function, Indicates the preset number of neighbor nodes, Indicates the The attention weights of neighbor nodes, Indicates the The attention projection matrix of neighbor nodes, Exponential function, represents the ReLU activation function with leakage, represents the attention score vector, represents the query matrix, Indicates the A key matrix, Indicates the A key matrix.

[0009] As an optional embodiment of the present invention, optionally, the expression of the cross entropy loss function is: ; in, represents the cross entropy loss function, batch size, Indicates the total number of fault categories, One-Hot encoding of the true label, Represents the sample predicted by the classifier Belong to category probability.

[0010] On the other hand, the present invention also provides a bearing intelligent fault diagnosis system based on generative adversarial network data enhancement and deep transfer learning, including the bearing intelligent fault diagnosis method based on generative adversarial network data enhancement and deep transfer learning; The system further comprises: Data acquisition module, used to collect vibration signals of bearings; A data enhancement module is used to perform data enhancement processing on the collected bearing vibration signal to generate an enhanced vibration signal; The feature extraction module is used to extract features from the enhanced vibration signal using a deep learning model to obtain deep time-frequency features; The transfer learning module is used to transfer deep time-frequency features from the source domain to the target domain based on the deep transfer learning strategy to obtain the transferred deep time-frequency features; The fault diagnosis module is used to identify the bearing fault type using a classifier based on the migrated deep time-frequency features to obtain the fault diagnosis results.

[0011] The present invention demonstrates the following benefits: By constructing an assisted classification generative adversarial network (ACGAN) to generate synthetic vibration signals consistent with the distribution of real fault samples, the present invention effectively alleviates the high cost of collecting bearing lifecycle fault data and the significant imbalance between normal and faulty samples. In particular, for situations where data on complex fault modes is extremely scarce, the ACGAN can generate synthetic samples that conform to the true physical distribution, thus resolving the problem that traditional data augmentation methods (such as superimposing noise and time domain scaling) can only achieve limited dimensionality expansion and cannot generate synthetic samples that conform to the true physical distribution. Furthermore, the present invention utilizes a deep learning model combined with a residual attention mechanism to extract features from the enhanced vibration signal, obtaining deep time-frequency features that better capture key information in the signal. Furthermore, through a deep transfer learning strategy, including domain adaptation and parameter transfer, the present invention minimizes the difference in feature distribution between the source and target domains, achieves knowledge transfer, and improves the model's generalization ability under different operating conditions. Finally, the present invention designs a classifier consisting of a feature mapping layer, a dynamic weight allocation layer, and a decision output layer. A gating mechanism is used to generate channel weights, optimizing the decision boundary and significantly improving the accuracy of bearing fault type identification.

[0012] Additional aspects and advantages of the present invention will be set forth in part in the description which follows and, in part, will be obvious from the description which follows, or may be learned by practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0013] The above and / or additional aspects and advantages of the present invention will become apparent and readily understood from the following description of the embodiments with reference to the accompanying drawings, in which: Figure 1 This is a flow chart of a bearing intelligent fault diagnosis method based on generative adversarial network data enhancement and deep transfer learning in the present invention. DETAILED DESCRIPTION

[0014] The following describes embodiments of the present invention in detail. Examples of the embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are intended only to explain the present invention and are not to be construed as limiting the present invention.

[0015] Example 1 like Figure 1 As shown, a bearing intelligent fault diagnosis method based on generative adversarial network data enhancement and deep transfer learning, the method includes: S1. Data enhancement of bearing vibration signals is performed through generative adversarial networks to generate synthetic vibration signals consistent with the distribution of real fault samples. S2. Using a deep learning model to extract features from the enhanced vibration signal to obtain deep time-frequency features; It should be noted in step S2 that the deep learning model used in this embodiment is a convolutional neural network (CNN). This CNN model comprises multiple convolutional layers, pooling layers, and fully connected layers. Through layer-by-layer convolution and pooling operations, it extracts discriminative deep time-frequency features from the enhanced vibration signal. These features efficiently represent key information in the bearing vibration signal. Furthermore, the CNN model, with its powerful feature learning capabilities, can automatically learn and extract the most useful features from the raw signal, eliminating the need for manual feature selection and extraction, significantly improving the efficiency and accuracy of fault diagnosis.

[0016] S3. Based on a deep transfer learning strategy, the deep time-frequency features are transferred from the source domain to the target domain to obtain the transferred deep time-frequency features. The deep transfer learning strategy includes domain adaptation and parameter transfer. The domain adaptation achieves knowledge transfer by minimizing the difference in feature distribution between the source domain and the target domain. The parameter transfer adapts to the new domain data by freezing the underlying parameters of the deep learning model and fine-tuning the high-level parameters. S4. Using a classifier to identify the bearing fault type based on the shifted deep time-frequency features, and obtain a fault diagnosis result.

[0017] The principle behind this embodiment of the intelligent bearing fault diagnosis method based on generative adversarial network data augmentation and deep transfer learning is to enhance the data of bearing vibration signals by constructing a variant of the generative adversarial network (GAN)—the auxiliary classification generative adversarial network (ACGAN). The ACGAN not only generates realistic synthetic vibration signals but also classifies these signals into fault categories, effectively addressing the scarcity of actual fault samples. During the data augmentation phase, the ACGAN gradually optimizes the parameters of the generator through adversarial training of the generator and the discriminator, making the generated synthetic vibration signals indistinguishable from real fault samples in feature space. Furthermore, the discriminator portion of the ACGAN also includes an auxiliary classifier for predicting the fault category of the input vibration signal (whether real or synthetic), further enhancing the consistency between the generated signal and real fault samples.

[0018] During the feature extraction phase, deep learning models, specifically convolutional neural networks (CNNs), are used to process the enhanced vibration signals. Leveraging its powerful feature learning capabilities, CNNs are able to extract discriminative deep time-frequency features from complex vibration signals. Through layer-by-layer convolution and pooling operations, CNNs automatically capture key information from the signal, eliminating the need for manual feature selection and extraction, significantly improving the efficiency and accuracy of fault diagnosis.

[0019] During the transfer learning phase, a deep transfer learning strategy is used to transfer deep time-frequency features from the source domain to the target domain. This process minimizes the differences between the feature distributions of the source and target domains through domain adaptation techniques, enabling effective knowledge transfer. Furthermore, a parameter transfer strategy freezes the underlying parameters of the deep learning model and fine-tunes higher-level parameters, enabling the model to better adapt to data from the new domain. This transfer learning strategy not only improves the model's generalization capabilities across diverse workloads but also reduces its reliance on large amounts of labeled data.

[0020] Finally, during the fault diagnosis phase, a classifier is used to identify bearing fault types based on the migrated deep time-frequency features. The classifier employs a carefully designed architecture, comprising a feature mapping layer, a dynamic weight assignment layer, and a decision output layer. The feature mapping layer maps the migrated deep time-frequency features into a latent space; the dynamic weight assignment layer generates channel weights through a gating mechanism, dynamically weighting different feature channels; and the decision output layer predicts fault probability based on the weighted features. By optimizing the classifier's parameters and structure, the accuracy and robustness of fault diagnosis can be further improved.

[0021] As an optional embodiment of the present invention, optionally, in step S1, performing data enhancement on the bearing vibration signal by generating an adversarial network includes: S101. Construct an auxiliary classification generative adversarial network, wherein the auxiliary classification generative adversarial network includes a generator and a discriminator. The generator is a one-dimensional convolutional network structure, including multiple upsampling blocks, each of which is composed of a transposed convolution layer, a batch normalization layer, and an activation function, and is used to map the input random noise and category label into a pseudo vibration signal; the discriminator adopts a one-dimensional convolutional network structure, including multiple downsampling blocks, each of which is composed of a convolution layer, a batch normalization layer, and an activation function, and the output of the discriminator includes two branches, one branch is used to output the probability that the signal is true, and the other branch is used to output the category probability distribution of the signal; S102, training the auxiliary classification generative adversarial network using loss function optimization; S103: Compare the pseudo vibration signal generated by the trained auxiliary classification generative adversarial network with the real fault sample, and continuously adjust the parameters of the generator and the discriminator so that the generated pseudo vibration signal is close to the real fault sample in distribution.

[0022] It should be noted that in step S103, multiple evaluation metrics, such as mean square error (MSE), peak signal-to-noise ratio (PSNR), and structural similarity index (SSIM), are used during the comparison process to ensure that the generated pseudo-vibration signal is of comparable quality to the actual fault sample. These metrics not only measure the numerical difference in the signal but also consider the signal's structural information and perceptual quality, thereby comprehensively evaluating the effectiveness of data augmentation.

[0023] As an optional embodiment of the present invention, optionally, the expression of the loss function in step S102 is: ; ; ; in, represents the discriminator loss, Indicates the expected value, Represents the real data distribution The real samples sampled from , The discriminator believes that is the probability of the true sample, Represents the distribution from the latent space Random noise sampled in , Represents the discriminator to generate samples The output, Represents the generator according to Generated pseudo samples, represents the weight coefficient of the gradient penalty, Represents the real sample and generate samples Samples obtained by random interpolation between , Representation Discriminator exist The gradient at represents the square of the deviation of the gradient modulus from 1, Represents uniform distribution A random number, represents the generator loss.

[0024] As an optional embodiment of the present invention, optionally, in step S2, the expression for extracting features from the enhanced vibration signal using the deep learning model is: ; ; ; ; in, represents the deep time-frequency features, represents the feature extraction function combined with the residual attention mechanism, represents the enhanced vibration signal, represents the learnable parameters, represents the Sigmoid activation function, represents the number of modes in the variational mode decomposition, represents the dynamic attention weight, Indicates the eigenmode components, Indicates the Residual blocks of different scales, Indicates the The residual block parameters of scales, Indicates that the input signal is decomposed into Decompose into eigenmode components, represents the second layer of convolution kernel, represents the ReLU activation function, represents the first layer of convolution kernel, represents the first layer bias, represents the exponential function, represents the transpose of the attention score vector, represents the hyperbolic tangent activation function, represents the projection layer weight, represents the projection layer bias, Indicates the eigenmode components.

[0025] As an optional embodiment of the present invention, optionally, in step S3, based on a deep transfer learning strategy, the deep time-frequency features are transferred from the source domain to the target domain, and obtaining the transferred deep time-frequency features includes: S301, constructing feature spaces of the source domain and the target domain, and using a domain adaptation method to make the feature distributions of the source domain and the target domain consistent by minimizing the KL divergence or the maximum average difference between the feature distributions of the source domain and the target domain; It should be noted in step S301 that KL divergence and maximum mean difference are important metrics for measuring the difference between two probability distributions. In this embodiment, KL divergence is selected as the optimization objective. By iteratively updating the parameters of the deep learning model, the difference between the feature distributions of the source and target domains is gradually reduced, thereby achieving effective knowledge transfer. At the same time, to ensure that the transferred features remain discriminative, a discriminative loss function is introduced to constrain the transferred features to maintain good inter-class separability and intra-class compactness in the target domain.

[0026] S302: Freeze the underlying parameters of the deep learning model, retain the common features learned in the source domain, and fine-tune the high-level parameters of the deep learning model to adapt to the characteristics of the target domain data; It should be noted that in step S302, during the fine-tuning process, a mini-batch gradient descent algorithm is used, combined with a learning rate decay strategy, to gradually optimize high-level parameters. This ensures that the model adapts to the new domain data while avoiding overfitting. Furthermore, to further enhance the effectiveness of transfer learning, a transfer learning regularization term is introduced to constrain the variation of model parameters during the transfer process, maintaining the model's stability and generalization capabilities. This series of strategies enables the effective transfer of deep time-frequency features from the source domain to the target domain.

[0027] S303: Continuously adjust the parameters of the deep learning model by iteratively optimizing the deep learning model until the migrated deep time-frequency features are obtained.

[0028] It should be noted in step S303 that early stopping is employed during the iterative optimization process to prevent model overfitting. Specifically, when the loss function value on the validation set stops decreasing significantly over several consecutive iterations, training is terminated, and the currently optimized model parameters are used as the final result. Furthermore, to accelerate convergence and improve training efficiency, a momentum optimizer and adaptive learning rate adjustment strategies (such as the Adam optimizer) are employed. These strategies dynamically adjust the learning rate based on historical gradient information, enabling the model to find the optimal solution more quickly during training. This series of optimization measures ensures the effectiveness and efficiency of the transfer learning phase.

[0029] As an optional embodiment of the present invention, optionally, in step S301, the adaptive method is an additional loss function; The expression of the additional loss function is: ; ; ; ; ; in, Indicates additional losses, represents the weight coefficient of classification supervision loss, represents the classification supervision loss, represents the probability predicted by the deep learning model, represents the true category of the sample, represents the weight coefficient of distribution alignment loss, represents the distribution alignment loss, represents the deep time-frequency features of the source domain, Represents the deep time-frequency features of the target domain, represents the weight coefficient of the covariance statistic matching loss, represents the covariance statistic matching loss, represents the covariance matrix of the source domain features, represents the covariance matrix of the target domain features, represents the weight coefficient of the gradient penalty term, represents the gradient penalty term, represents the random interpolation of source and target domain features, represents the batch size, Indicates the number of fault categories, Indicates the Does the sample belong to kind, Represents the first The samples belong to The probability of the class, represents the number of samples in the batch, represents the discriminator network, Indicates the source domain The deep time-frequency features of samples, Indicates the target domain The deep time-frequency features of samples, Represents the dimension of deep time-frequency features, Indicates the expected value, Representation Discriminator exist The gradient at .

[0030] As an optional embodiment of the present invention, optionally, in step S4, identifying the bearing fault type using a classifier based on the shifted deep time-frequency features to obtain a fault diagnosis result includes: S401. Construct a classifier, which includes a feature mapping layer, a dynamic weight allocation layer, and a decision output layer. The feature mapping layer adopts a fully connected structure to map the migrated deep time-frequency features to a latent space; the dynamic weight allocation layer generates channel weights through a gating mechanism; and the decision output layer predicts fault probability based on weighted features. It should be noted in step S401 that the main function of the feature mapping layer is to map the migrated deep time-frequency features from the original feature space to a new latent space for better classification. Through the fully connected structure, the complex relationship between features can be learned, thereby extracting feature representations that are more favorable for classification tasks. The dynamic weight allocation layer dynamically allocates weights according to the importance of each feature by introducing a gating mechanism. This mechanism can adaptively adjust the impact of different features on the final classification results, thereby improving the accuracy of classification. The decision output layer predicts the fault probability based on the weighted features, and ultimately determines the fault type of the bearing by calculating the probability of each fault category. This series of operations ensures that the bearing fault type can be accurately identified based on the migrated deep time-frequency features, thereby improving the accuracy and reliability of fault diagnosis.

[0031] S402, training the classifier using a cross entropy loss function, and continuously adjusting the parameters of the classifier to minimize the difference between the true label and the predicted label; It should be noted in step S401 that the cross-entropy loss function, a commonly used loss function in classification tasks, can effectively measure the difference between the true label and the predicted label. By minimizing this loss function, the classifier parameters can be continuously adjusted to better fit the training data. In this embodiment, the classifier is trained using a mini-batch gradient descent algorithm, combined with a learning rate decay strategy to ensure rapid model convergence during training while avoiding overfitting. This training strategy ensures that the classifier can accurately identify the type of bearing fault.

[0032] S403. During the training process, an early stopping strategy is adopted. When the loss on the validation set no longer decreases, the training is stopped and the optimal classifier is saved. It should be noted in step S403 that the early stopping strategy is a regularization method that can prevent model overfitting during training. When the loss function value on the validation set no longer decreases significantly over several consecutive iterations, the model is considered to have converged to the optimal solution. At this point, training is stopped and the current optimal classifier model is saved. This prevents performance degradation due to overfitting in subsequent iterations. This strategy ensures that the resulting classifier model has both good generalization capabilities and high accuracy in bearing fault type identification tasks.

[0033] S404: Input the vibration signal of the test set into the trained classifier and output the fault diagnosis result.

[0034] It should be noted in step S404 that after the vibration signal of the test set is input into the classifier, the classifier will perform feature extraction, transfer learning and classification prediction on the input signal based on the pre-learned knowledge. Specifically, the classifier will first pre-process the vibration signal of the test set to extract the key features related to the bearing fault. Then, using the deep transfer learning strategy, these features are migrated from the source domain (i.e., the domain where the training data is located) to the target domain (i.e., the domain where the test data is located) to adapt to the fault diagnosis needs under different working conditions. Finally, based on the migrated features, the classifier will predict the fault probability and output the final fault diagnosis result. This result will directly indicate the fault type of the bearing. The above-mentioned method of the present invention can realize intelligent, fast and accurate diagnosis of bearing faults, greatly improving the efficiency and accuracy of fault diagnosis.

[0035] As an optional embodiment of the present invention, optionally, the expression of the classifier is: ; ; ; in, represents the failure probability vector output by the classifier, Represents the deep time-frequency features after the fault classifier is transferred Mapped to a failure probability vector The composite operation, represents the deep time-frequency features after migration, represents the set of learnable parameters of the classifier, represents the Softmax function, represents the projection matrix, represents the graph attention weight vector, represents the fully connected projection matrix, represents the bias vector, represents the Sigmoid activation function, Indicates the preset number of neighbor nodes, Indicates the The attention weights of neighbor nodes, Indicates the The attention projection matrix of neighbor nodes, Exponential function, represents the ReLU activation function with leakage, represents the attention score vector, represents the query matrix, Indicates the A key matrix, Indicates the A key matrix.

[0036] As an optional embodiment of the present invention, optionally, the expression of the cross entropy loss function is: ; in, represents the cross entropy loss function, batch size, Indicates the total number of fault categories, One-Hot encoding of the true label, Represents the sample predicted by the classifier Belong to category probability.

[0037] Example 2 A bearing intelligent fault diagnosis system based on generative adversarial network data enhancement and deep transfer learning, including the bearing intelligent fault diagnosis method based on generative adversarial network data enhancement and deep transfer learning; The system further comprises: Data acquisition module, used to collect vibration signals of bearings; A data enhancement module is used to perform data enhancement processing on the collected bearing vibration signal to generate an enhanced vibration signal; The feature extraction module is used to extract features from the enhanced vibration signal using a deep learning model to obtain deep time-frequency features; The transfer learning module is used to transfer deep time-frequency features from the source domain to the target domain based on the deep transfer learning strategy to obtain the transferred deep time-frequency features; The fault diagnosis module is used to identify the bearing fault type using a classifier based on the migrated deep time-frequency features to obtain the fault diagnosis results.

[0038] The bearing intelligent fault diagnosis system based on generative adversarial network data enhancement and deep transfer learning in this embodiment is used to implement the bearing intelligent fault diagnosis method based on generative adversarial network data enhancement and deep transfer learning in Example 1.

[0039] In this embodiment, the intelligent bearing fault diagnosis system based on generative adversarial network data augmentation and deep transfer learning works as follows: The system first collects bearing vibration signals through the data acquisition module. The data augmentation module then processes the original vibration signals using techniques such as generative adversarial networks to generate more diverse enhanced vibration signals, thereby improving the model's generalization capabilities. The feature extraction module uses deep learning models, such as convolutional neural networks or recurrent neural networks, to extract features from the enhanced vibration signals, extracting deep time-frequency features that are sensitive and stable to bearing faults. The transfer learning module then employs a deep transfer learning strategy to transfer these features from the source domain (i.e., the environment or conditions in which the training data was generated) to the target domain (i.e., the environment or conditions in which the actual diagnostic data was generated), adapting to the fault diagnosis requirements under different operating conditions. Finally, the fault diagnosis module uses a trained classifier based on the transferred deep time-frequency features to identify the bearing fault type and output a fault diagnosis result. This result provides a key basis for subsequent maintenance decisions. Through the collaborative work of these modules, the system can achieve intelligent, rapid, and accurate diagnosis of bearing faults, significantly improving the efficiency and accuracy of fault diagnosis.

[0040] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and alterations may be made to the embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the claims and their equivalents.

Claims

1. A bearing intelligent fault diagnosis method based on generative adversarial network data enhancement and deep transfer learning, characterized in that: The method comprises: S1. Data enhancement of bearing vibration signals is performed through generative adversarial networks to generate synthetic vibration signals consistent with the distribution of real fault samples. S2. Using a deep learning model to extract features from the enhanced vibration signal to obtain deep time-frequency features; S3. Based on a deep transfer learning strategy, the deep time-frequency features are transferred from the source domain to the target domain to obtain the transferred deep time-frequency features. The deep transfer learning strategy includes domain adaptation and parameter transfer. The domain adaptation achieves knowledge transfer by minimizing the difference in feature distribution between the source domain and the target domain. The parameter transfer adapts to the new domain data by freezing the underlying parameters of the deep learning model and fine-tuning the high-level parameters. S4. Using a classifier to identify the bearing fault type based on the shifted deep time-frequency features, and obtain a fault diagnosis result.

2. The bearing intelligent fault diagnosis method based on generative adversarial network data enhancement and deep transfer learning according to claim 1 is characterized in that: In step S1, data enhancement of the bearing vibration signal by generating an adversarial network includes: S101. Construct an auxiliary classification generative adversarial network, wherein the auxiliary classification generative adversarial network includes a generator and a discriminator. The generator is a one-dimensional convolutional network structure, including multiple upsampling blocks, each of which is composed of a transposed convolution layer, a batch normalization layer, and an activation function, and is used to map the input random noise and category label into a pseudo vibration signal; the discriminator adopts a one-dimensional convolutional network structure, including multiple downsampling blocks, each of which is composed of a convolution layer, a batch normalization layer, and an activation function, and the output of the discriminator includes two branches, one branch is used to output the probability that the signal is true, and the other branch is used to output the category probability distribution of the signal; S102, training the auxiliary classification generative adversarial network using loss function optimization; S103: Compare the pseudo vibration signal generated by the trained auxiliary classification generative adversarial network with the real fault sample, and continuously adjust the parameters of the generator and the discriminator so that the generated pseudo vibration signal is close to the real fault sample in distribution.

3. The bearing intelligent fault diagnosis method based on generative adversarial network data enhancement and deep transfer learning according to claim 2 is characterized in that: The loss function in step S102 is expressed as: ； ； ； in, represents the discriminator loss, Indicates the expected value, Represents the real data distribution The real samples sampled from , The discriminator believes that is the probability of the true sample, Represents the distribution from the latent space Random noise sampled in , Represents the discriminator to generate samples The output, Represents the generator according to Generated pseudo samples, represents the weight coefficient of the gradient penalty, Represents the real sample and generate samples Samples obtained by random interpolation between , Representation Discriminator exist The gradient at represents the square of the deviation of the gradient modulus from 1, Represents uniform distribution A random number, represents the generator loss.

4. The bearing intelligent fault diagnosis method based on generative adversarial network data enhancement and deep transfer learning according to claim 1 is characterized in that: In step S2, the expression for extracting features from the enhanced vibration signal using the deep learning model is: ； ； ； ； in, represents the deep time-frequency features, represents the feature extraction function combined with the residual attention mechanism, represents the enhanced vibration signal, represents the learnable parameters, represents the Sigmoid activation function, represents the number of modes in the variational mode decomposition, represents the dynamic attention weight, Indicates the eigenmode components, Indicates the Residual blocks of different scales, Indicates the The residual block parameters of scales, Indicates that the input signal is decomposed into Decompose into eigenmode components, represents the second layer of convolution kernel, represents the ReLU activation function, represents the first layer of convolution kernel, represents the first layer bias, represents the exponential function, represents the transpose of the attention score vector, represents the hyperbolic tangent activation function, represents the projection layer weight, represents the projection layer bias, Indicates the eigenmode components.

5. The bearing intelligent fault diagnosis method based on generative adversarial network data enhancement and deep transfer learning according to claim 1 is characterized in that: In step S3, based on the deep transfer learning strategy, the deep time-frequency features are transferred from the source domain to the target domain, and the transferred deep time-frequency features are obtained, including: S301, constructing feature spaces of the source domain and the target domain, and using a domain adaptation method to make the feature distributions of the source domain and the target domain consistent by minimizing the KL divergence or the maximum average difference between the feature distributions of the source domain and the target domain; S302: Freeze the underlying parameters of the deep learning model, retain the common features learned in the source domain, and fine-tune the high-level parameters of the deep learning model to adapt to the characteristics of the target domain data; S303: Continuously adjust the parameters of the deep learning model by iteratively optimizing the deep learning model until the migrated deep time-frequency features are obtained.

6. The bearing intelligent fault diagnosis method based on generative adversarial network data enhancement and deep transfer learning according to claim 5 is characterized in that: In step S301, the adaptive method is an additional loss function; The expression of the additional loss function is: ； ； ； ； ； in, Indicates additional losses, represents the weight coefficient of classification supervision loss, represents the classification supervision loss, represents the probability predicted by the deep learning model, represents the true category of the sample, represents the weight coefficient of distribution alignment loss, represents the distribution alignment loss, represents the deep time-frequency features of the source domain, Represents the deep time-frequency features of the target domain, represents the weight coefficient of the covariance statistic matching loss, represents the covariance statistic matching loss, represents the covariance matrix of the source domain features, represents the covariance matrix of the target domain features, represents the weight coefficient of the gradient penalty term, represents the gradient penalty term, represents the random interpolation of source and target domain features, represents the batch size, Indicates the number of fault categories, Indicates the Does the sample belong to kind, Represents the first The samples belong to The probability of the class, represents the number of samples in the batch, represents the discriminator network, Indicates the source domain The deep time-frequency features of samples, Indicates the target domain The deep time-frequency features of samples, Represents the dimension of deep time-frequency features, Indicates the expected value, Representation Discriminator exist The gradient at .

7. The bearing intelligent fault diagnosis method based on generative adversarial network data enhancement and deep transfer learning according to claim 1 is characterized in that: In step S4, the bearing fault type is identified using a classifier based on the shifted deep time-frequency features to obtain a fault diagnosis result, including: S401. Construct a classifier, which includes a feature mapping layer, a dynamic weight allocation layer, and a decision output layer. The feature mapping layer adopts a fully connected structure to map the migrated deep time-frequency features to a latent space; the dynamic weight allocation layer generates channel weights through a gating mechanism; and the decision output layer predicts fault probability based on weighted features. S402, training the classifier using a cross entropy loss function, and continuously adjusting the parameters of the classifier to minimize the difference between the true label and the predicted label; S403. During the training process, an early stopping strategy is adopted. When the loss on the validation set no longer decreases, the training is stopped and the optimal classifier is saved. S404: Input the vibration signal of the test set into the trained classifier and output the fault diagnosis result.

8. The bearing intelligent fault diagnosis method based on generative adversarial network data enhancement and deep transfer learning according to claim 1 or 7, characterized in that: The expression of the classifier is: ； ； ； in, represents the failure probability vector output by the classifier, Represents the deep time-frequency features after the fault classifier is transferred Mapped to a failure probability vector The composite operation, represents the deep time-frequency features after migration, represents the set of learnable parameters of the classifier, represents the Softmax function, represents the projection matrix, represents the graph attention weight vector, represents the fully connected projection matrix, represents the bias vector, represents the Sigmoid activation function, Indicates the preset number of neighbor nodes, Indicates the The attention weights of neighbor nodes, Indicates the The attention projection matrix of neighbor nodes, Exponential function, represents the ReLU activation function with leakage, represents the attention score vector, represents the query matrix, Indicates the A key matrix, Indicates the A key matrix.

9. The bearing intelligent fault diagnosis method based on generative adversarial network data enhancement and deep transfer learning according to claim 7 is characterized in that: The expression of the cross entropy loss function is: ； in, represents the cross entropy loss function, batch size, Indicates the total number of fault categories, One-Hot encoding of the true label, Represents the sample predicted by the classifier Belong to category probability.

10. A bearing intelligent fault diagnosis system based on generative adversarial network data enhancement and deep transfer learning, characterized in that: Including the bearing intelligent fault diagnosis method based on generative adversarial network data enhancement and deep transfer learning as described in any one of claims 1 to 9; The system further comprises: Data acquisition module, used to collect vibration signals of bearings; A data enhancement module is used to perform data enhancement processing on the collected bearing vibration signal to generate an enhanced vibration signal; The feature extraction module is used to extract features from the enhanced vibration signal using a deep learning model to obtain deep time-frequency features; The transfer learning module is used to transfer deep time-frequency features from the source domain to the target domain based on the deep transfer learning strategy to obtain the transferred deep time-frequency features; The fault diagnosis module is used to identify the bearing fault type using a classifier based on the migrated deep time-frequency features to obtain the fault diagnosis results.

Images